…(v3.1)

- Port upstream _causal_conv1d_fwd_kernel as NPU Triton kernel
  - Handles initial/final/intermediate conv state in-kernel
  - Supports APC block boundary state writes
  - NPU adaptations: removed .cache_modifier, kept debug_barrier
- Rewrite causal_conv1d_fn to dispatch to new Triton kernel
- Rewrite gdn.py conv1d path: split decode/prefill like upstream
  - Decode: causal_conv1d_update_npu with block params
  - Prefill: causal_conv1d_fn with APC params (new kernel)
- Fix SSM #6: _build_initial_state only zeros prefill sequences
- Fix SSM #7: _write_final_states adds slot >= 0 validation
- Fix SSM #8: _scatter_intermediate_states adds unaligned offset
- Update all 36 UTs to pass with new num_computed_tokens_all field

Alignment status vs upstream #26807:
  #1 conv1d prefill kernel:     FIXED (kernel ported)
  #3 causal_conv1d_fn params:   FIXED (rewritten)
  #4 intermediate conv state:   FIXED (kernel internal)
  #6 SSM zeroing scope:         FIXED
  #7 _write_final_states guard: FIXED
  #8 SSM scatter alignment:     FIXED
  #9 causal_conv1d_fn signature: FIXED
  #2 decode pre-copy:           KEEP (NPU needs it)
  #5 SSM decode index:          OK (correct approach)
  #10 conv layout hardcoded:    DEFERRED

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Five logic / hot-path issues in vllm_ascend/ops/triton/mamba/causal_conv1d.py:

1. (#1) Fallback non-APC `causal_conv1d_fn` ignored `has_initial_state[i]` and
   always seeded the convolution from `conv_states[cache_indices[i]]`. First
   prefill iterations (no prior context) silently used stale cache state.
   Now passes `initial_states=None` when `has_initial_state[i]` is False.

2. (#2) `forward_context` lookup at the top of `causal_conv1d_fn` ran on
   every per-layer APC prefill call, but `num_decodes` is only used by the
   non-APC fallback path. Moved the lookup inside the fallback branch so the
   APC hot path does not pay it 28x per prefill on Qwen3.5.

3. (#3) `NULL_BLOCK_ID` fell back to `PAD_SLOT_ID` (-1) when upstream did
   not export it. vLLM v1 `BlockPool` reserves block_id=0 as the null block
   and vLLM-Ascend `block_table.fill_(0)` matches that semantics, while
   PAD_SLOT_ID=-1 only applies to 1D slot/state-index padding. The fallback
   should be 0 so the `null_block_id != pad_slot_id` normalization actually
   masks 2D block-table placeholders. Documented the distinction at the
   constant definition site.

4. (#4) `_normalize_apc_input_layout` did four heuristic shape comparisons
   to auto-detect dim-major / weight-transposed inputs. This is ambiguous
   when dim==width or dim==tokens, and shifts layout-fix cost into the
   prefill hot path. Replaced with an explicit contract: x is [tokens, dim],
   weight is [dim, width]; fail loud otherwise. Removed the now-unused
   `input_is_dim_major` plumbing and the trailing `out.transpose(0,1).
   contiguous()` in `_causal_conv1d_fwd_npu`.

5. (#5) `_uses_apc_prefill_path` selected APC when ANY of five tensors was
   non-None, masking caller bugs (e.g. block_idx_* provided without a 2D
   block table). Now selects on cache_indices.dim()==2 and raises
   RuntimeError if APC auxiliary tensors are passed without a 2D block
   table, matching the documented dispatcher contract.

Also dropped the redundant `conv_states.shape[-2] != dim and shape[-1] == dim:
transpose` probe in the dispatcher; the caller already passes the standard
`[num_cache_lines, dim, state_len]` layout, and silent re-permutation hides
upstream contract violations.

No behaviour change on the production caller path (gdn._run_all_mode_non_spec_conv1d),
verified by tracing layout invariants. The fixes harden the dispatcher
against future caller refactors and remove dead/wrong code.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Signed-off-by: lHrHenry233 <2381623149@qq.com>